Steer ViT-S away from ViT-B to bracket cross-model similarity

Construct a function-preserving HNC walk that steers an ImageNet-trained ViT-S away from a frozen ViT-B, thereby bracketing the similarity between the two trained models from below as well as from above.

Background

The paper demonstrates that a ViT-S can be steered toward a frozen ViT-B while largely preserving the ViT-S predictions, showing that cross-model representational similarity can be increased through a function-preserving HNC trajectory. The complementary experiment—steering ViT-S away from ViT-B—would provide the corresponding lower bracket on the similarity between independently trained models. This would more fully characterize the range of representational similarity accessible at approximately fixed function.

References

The complementary walk, steering ViT-S away from ViT-B, would bracket the similarity of two trained models from below as well as from above, and we leave it to future work.

— Traversing the solution space of neural networks with Hessian Null Space Continuation  (2609.38081 - Huang et al., 29 Sep 2026) in Appendix, Section “Steered convergence between ViT-S and ViT-B” (Appendix \ref{app:prh_vits_vitb})